When to Diverge
Diverge when a single viewpoint isn't enough:
- - Product decisions — Different user types have conflicting needs
- Safety-critical — Need doctor, lawyer, security expert perspectives
- Creative work — Multiple aesthetic directions before choosing
- Complex trade-offs — No obvious "right" answer
Don't diverge for:
- - Simple tasks with clear answers
- Speed-critical requests
- Tasks where you already have high confidence
How Many Perspectives
| Complexity | Perspectives | Examples |
|---|
| Low | 2-3 | Quick sanity check |
| Medium |
3-5 | Product feature review |
| High | 5-7 | Safety-critical, legal |
Rule: Enough to cover blind spots, not so many you can't synthesize.
Choosing Perspectives
Match perspectives to the problem:
- - Product: Power user, casual user, churned user, new user
- Medical: Doctor, patient, caregiver, insurance
- Legal: Lawyer, affected party, regulator
- Technical: Security, performance, maintainability
- Content: Target audience, critic, competitor
Check personas.md for common perspective templates.
Synthesis
After collecting perspectives:
- 1. Identify conflicts — Where do perspectives disagree?
- Find common ground — What do all perspectives agree on?
- Weigh by stakes — Safety concerns > preferences
- Decide — Make a call, document trade-offs
Check synthesis.md for conflict resolution patterns.
Learning User Preferences
Track when divergence helped vs. was overkill. Adapt:
- - User who values speed → diverge less, ask before spawning
- User who values thoroughness → diverge more proactively
Related: For routing tasks to sub-agents, see
delegate. For iterating until success, see
loop.
何时发散
当单一视角不足时，请发散：
- - 产品决策 — 不同用户类型存在冲突需求
- 安全关键 — 需要医生、律师、安全专家视角
- 创意工作 — 在选定前探索多种美学方向
- 复杂权衡 — 没有明显的正确答案
以下情况无需发散：
- - 有明确答案的简单任务
- 对速度要求高的请求
- 你已高度确信的任务
所需视角数量
3-5 | 产品功能评审 |
| 高 | 5-7 | 安全关键、法律事务 |
原则： 足够覆盖盲区即可，不宜过多导致无法综合。
选择视角
将视角与问题匹配：
- - 产品： 核心用户、普通用户、流失用户、新用户
- 医疗： 医生、患者、护理者、保险公司
- 法律： 律师、受影响方、监管机构
- 技术： 安全、性能、可维护性
- 内容： 目标受众、评论家、竞争对手
查看 personas.md 获取常见视角模板。
综合处理
收集视角后：
- 1. 识别冲突 — 各视角在哪些方面存在分歧？
- 寻找共识 — 所有视角都认同什么？
- 按重要性权衡 — 安全问题优先于偏好
- 做出决策 — 拍板定案，记录权衡取舍
查看 synthesis.md 获取冲突解决模式。
学习用户偏好
追踪发散何时有效、何时过度。根据情况调整：
- - 重视速度的用户 → 减少发散，生成前先询问
- 重视全面性的用户 → 更主动地发散
相关： 如需将任务路由给子代理，请参见 delegate。如需迭代直至成功，请参见 loop。
